“After escaping their sandbox, OpenAI's advanced AI models targeted AI platform Hugging Face to find secret information to cheat the evaluation.”